Center Goals

• Conduct research leading to the design and construction of software tools and systems to help applications scale to the petascale and beyond
• Focus on DOE Leadership Class Facilities and parallel systems composed of multicore processors
• Promote application-driven software systems research
• Promote research collaborations with DOE (Cyberinfrastructure), and industry (systems and software vendors)
  

• Catalyze activities within the computer science community that will lead to visionary new ideas for application development support software
• Focus on interactions with systems vendors, application developers, and library designers
• Promote community vision building through summer workshops
• Foster development of new tools by the computer science community through support of common software infrastructures and standards
• Focus on open-source software development and integration, including HPCToolkit
  

Scalable Application Development

• Scaling from High-level to High-Performance
• Example: Translate Matlab to C or Fortran
• Constraint-based type analysis: determine array sizes and types
• Replace calls to domain library with calls to pre-optimized versions, specialized to parameter type signature
• Scalarize array statements
• Example: Source-to-source performance optimization
• Inlining, loop fusion, tiling for multiple levels of cache
• Scaling from One to Many Processors
• Example: Matlab with data distributions
• Direct translation to Fortran+MPI via HPF compilation strategy
• Example: Tools for performance understanding
• HPCToolkit reveals where parallel efficiency is lost
• Scaling to One to Many Platforms
• Example: Automatic tuning to different machine architectures

Research Focus Areas

• Compilers, libraries, and tuning systems
• Automatic tuning for multicore chips and scalable systems
• Programming models and compilers for scalable parallel computing
• Co-Array Fortran, UPC, Titanium, HPCS languages
• Extensions to standard languages (C, C++, Fortran)
• High-level domain languages based on scripting languages
• Example: Matlab (with parallelism) plus domain libraries
• High performance component integration frameworks
• Application of telescoping languages
• Performance tools for scalable parallel systems
• Binary analysis for instrumentation and interpretation of measurements